Method and system for extracting chapter-level events based on multi-round machine reading understanding

Abstract

The invention discloses a chapter-level event extraction method and system based on multi-round machine reading understanding, and relates to the technical field of event extraction methods, and the method comprises the following steps: S1, building a question template; s2, splicing an input sequence; s3, generating an embedded representation; s4, answer matching and positioning; s5, model training optimization; and S6, dynamic threshold reasoning. According to the method, the argument roles with similar semantics in different event types are associated through the interrogative words with unified semantic attributes through a method of classifying according to the argument role types, matching the corresponding interrogative words and generating the multi-class question templates, a semantic sharing channel of cross-event types is established, the limitation of independent classification of similar roles in a traditional method is broken through, and the method has the advantages of being high in efficiency and high in efficiency. The model can efficiently reuse semantic features of similar roles, the recognition generalization ability of similar argument roles is remarkably improved, and the argument extraction requirements under multiple event types are met.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of event extraction methods, in particular to a chapter-level event extraction method and system based on multi-round machine reading comprehension. BACKGROUND

[0002] In the field of natural language processing, chapter-level event extraction, as a core task of accurately identifying event types, locating event triggers (i.e. core words triggering events) and extracting event arguments (such as event participants, occurrence time, involved locations, associated equipment, and other key elements) from long texts such as news reports and research papers, plays an important role in converting unstructured text into structured event knowledge in key scenarios such as military news situation analysis, public health emergency event monitoring, and national security intelligence information mining, and has high application value and practical significance.

[0003] However, the current traditional event extraction method still faces significant technical bottlenecks when adapting to chapter-level text processing needs. On the one hand, the semantic features of similar argument roles in different event types are not effectively shared, and the model's generalization ability for similar argument roles is greatly limited, especially in the context of domain-specific argument roles in military and emergency fields, where performance degradation is more pronounced.

[0004] On the other hand, traditional event extraction methods generally use a pipeline process of trigger word detection-event type classification-argument extraction. This process has inherent error propagation defects. If errors occur in the previous steps (such as trigger word recognition), the errors will be directly transmitted to the subsequent argument extraction step, resulting in a continuous decline in overall extraction accuracy. Moreover, such methods generally use fixed thresholds to filter candidate results, which cannot dynamically adjust the filtering criteria according to the context differences of chapter texts (such as sentence complexity and argument distribution density). When faced with complex scenarios such as multiple event overlaps (a text containing multiple related events), missing arguments (some event elements not explicitly stated in the text), and cross-sentence argument distribution (arguments of the same event scattered in different paragraphs), fixed threshold filtering may miss effective arguments or mistakenly select irrelevant fragments, making it difficult to balance the recall and precision of event extraction, and unable to meet the requirements of extraction accuracy in practical applications. SUMMARY

[0005] The present application aims to provide a chapter-level event extraction method and system based on multi-round machine reading comprehension to solve the problems raised in the background.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a method for extracting passage-level events based on multi-turn machine reading comprehension, comprising the following steps:

[0007] S1. Question Template Construction: For the three sub-tasks of the chapter-level event extraction task, namely event trigger word identification, trigger word classification and event argument extraction, an appropriate question template is designed for each. The question template incorporates the semantic information of argument roles and event association features, and the event argument extraction template matches the corresponding question words according to the argument role type.

[0008] S2. Input sequence concatenation: The question template constructed in S1 is concatenated with the text to be processed according to a preset format to form a model input sequence. The preset format includes a start indicator, a separator, and a fixed arrangement order of questions and text.

[0009] S3, Embedded Representation Generation: The input sequence obtained in S2 is encoded using a pre-trained language model to generate an initial semantic embedding vector; then, the initial semantic embedding vector is processed by a bidirectional gated recurrent unit to capture the bidirectional dependencies of the text and output the final embedded representation.

[0010] S4. Answer matching and localization: Using two independent binary classifiers, the probability of each tag in the final embedded representation generated in S3 as the start and end positions of the event trigger word, event type, and event argument is predicted to obtain the probability distribution of the candidate answer position.

[0011] S5. Model Training and Optimization: Based on the candidate answer position probability distribution and true label obtained in S4, the cross-entropy loss function is used to calculate the loss value, the optimizer is used to minimize the loss value, and the model parameters are updated.

[0012] S6. Dynamic Threshold Reasoning: Determine the optimal probability threshold based on the validation set, calculate the no-answer score of the candidate span by combining the candidate answer position probability distribution obtained in S4, and select the candidate spans whose no-answer scores meet the threshold conditions as the final event trigger words and event arguments to complete the chapter-level event extraction.

[0013] Preferably, the specific process of constructing the problem template in step S1 includes:

[0014] S11. Construction of Trigger Word Recognition Template: Design a general interrogative sentence template for locating potential event trigger words. The template is composed of interrogative words combined with the core expression of trigger words, and is used to guide the model to filter out continuous words that can be used as event trigger words from the text.

[0015] S12. Construction of Trigger Word Classification Template: Design an interrogative template containing the located potential trigger words, the location information of the trigger words in the text, and the candidate event types. The template is constructed with a confirmatory expression to guide the model to judge the matching relationship between potential trigger words and candidate event types, and then determine the event type corresponding to the trigger words.

[0016] S13. Event Argument Extraction Template Construction: First, classify event argument roles into categories such as event participants, event location attributes, event time attributes, and event equipment, scale, or content attributes. For different categories of argument roles, match corresponding interrogative words: if the argument role is an event participant, select interrogative words pointing to the subject; if it is an event location attribute, select interrogative words pointing to the spatial location; if it is an event time attribute, select interrogative words pointing to the time range; if it is an event equipment, scale, or content attribute, select interrogative words pointing to the object or state. Based on the matched interrogative words and argument role names, generate three types of event argument extraction templates: the first type is a template that only contains the argument role name; the second type is a template that contains the corresponding interrogative words and the argument role name; and the third type is a template manually generated in the form of natural language, combined with the context of the text, to adapt to the argument extraction needs of different semantic complexities.

[0017] Preferably, the specific method for concatenating the input sequence in step S2 is as follows: according to The pieces are assembled in the following order, among which... As the start indicator, Used as a separator; Let m be a sequence of problems. , The i-th word in the question; Let n be a sequence of text passages of length n. , Let j be the j-th word in the text; the total length of the concatenated input sequence is n+m+3 tokens.

[0018] Preferably, the specific process of embedding characterization generation in step S3 includes:

[0019] S31. Pre-trained language model encoding: Input the concatenated input sequence from S2 into the pre-trained language model. Output the initial semantic embedding vector , Where l is the length of the input sequence, and d is... The hidden layer dimension is R, which is the set of real numbers and satisfies the formula. (For illustration) Source: By concatenating the input sequence Send in The model is encoded, and the final output is the initial semantic embedding vector. );

[0020] S32, Processing: The initial semantic embedding vector obtained in S31 Input to bidirectional gated loop unit By employing a bidirectional gating mechanism to capture the positive and negative dependencies in the text, an enhanced contextual feature vector C is output. ,in This is the enhanced feature vector corresponding to the i-th label in the input sequence;

[0021] S33. Probability Distribution Calculation: Perform linear transformations on the context feature vector C obtained in S32, and introduce the initial position weight matrix. End position weight matrix and the corresponding bias terms , Then, the probability of each marker being the start and end position is calculated using the Softmax function, satisfying the formula. , ,in Let be the probability of using the i-th marker as the starting position. Let the probability of the i-th mark being the ending position be given; thus, the probability distribution of the starting position can be obtained. Probability distribution of the ending position And the probability distribution of the starting position Probability distribution of the ending position Used for subsequent answer matching and dynamic threshold reasoning steps.

[0022] Preferably, the specific process of answer matching and location in step S4 includes:

[0023] S41. Preliminary trigger word localization: based on the probability distribution of the starting position output from S3. Probability distribution of the ending position The continuous marker sequences with probability values ​​greater than the preset initial threshold are selected as potential event trigger word candidates for subsequent screening by the dynamic threshold reasoning module.

[0024] S42. Preliminary Event Type Matching: Substitute the potential event trigger word candidates obtained in S41 and their positions in the text into the trigger word classification template constructed in S1, and combine... and Calculate the matching probability between the trigger word candidate and each candidate event type, select the candidate event type with the highest matching probability as the event type corresponding to the trigger word candidate, and form a trigger word candidate-event type association pair for subsequent verification by the dynamic threshold inference module;

[0025] S43. Preliminary Argument Localization: Based on the trigger word candidate-event type association pair formed in S42, the event argument extraction template for the corresponding event type constructed in S1 is called. According to the interrogative words and argument role information in the template, from... and Continuous marker sequences that match semantic roles are selected as event argument candidates for final selection by the dynamic threshold inference module.

[0026] Preferably, the specific process of model training optimization in step S5 includes:

[0027] S51. Loss Calculation: Calculate the initial position loss using the cross-entropy loss function. Loss at the end position Satisfying the formula , ,in The starting position is the actual label sequence. Set the actual label sequence at the end position; set hyperparameters ∈[0,1], calculate the total loss And satisfy the formula ;

[0028] S52, Parameter Update: Using the AdamW optimizer, the total loss is calculated based on S51. The model training is optimized by updating the parameters of the pre-trained language model, bidirectional gated recurrent unit, and binary classifier in the model through backpropagation.

[0029] Preferably, the parameter configuration and execution steps for model training include:

[0030] S81. Set the hidden layer dimension: Initialize the hidden layer dimension of the pre-trained language model and the bidirectional gated recurrent unit, and set it to a dimension value with sufficient semantic representation ability to ensure that the model can capture complex semantic relationships in the text.

[0031] S82. Configure batch processing parameters: First, set a reasonable batch size, and then adopt the cumulative step size strategy, that is, calculate the gradient for a preset number of consecutive small batches of training data, accumulate the gradients and perform a parameter update operation. This method simulates the effect of large batch training, reduces gradient oscillation in small batch training, and improves the stability of model training.

[0032] S83. Determine the number of training epochs and the optimizer: Set the total number of training epochs to the number of iterations that ensures the model can converge to the optimal parameter state; select AdamW as the model optimizer, and use the characteristic of the optimizer supporting weight decay to regularize the model parameters and suppress model overfitting.

[0033] S84. Set learning rate and text length limits: Set the learning rate of the model training to a value that balances the learning speed and the stability of parameter updates, so as to avoid the model training from not converging due to the learning rate being too high or the training efficiency being low due to the learning rate being too low; at the same time, limit the maximum length of the input text, so as to ensure that the model covers enough text information to handle long texts of the chapter level, while controlling the computational complexity of the model and ensuring that the training process proceeds efficiently.

[0034] Preferably, the specific process of dynamic threshold inference in step S6 includes:

[0035] S61. Argument Span Collection: Traversing all possible starting positions in the text. and end position ,in , ∈{1,...,M}, where M is the text length, filter spans that satisfy the following conditions: and Within the range of M, , and ,in , These are the start indicators. The corresponding starting and ending position probabilities; calculate the score for the span after filtering. Score for no answer ,Will Store in a candidate list, which includes potential event trigger word candidates in S41, trigger word candidate-event type association pairs in S42, and span information corresponding to event argument candidates in S43;

[0036] S62. Determining the optimal threshold: Based on the candidate list corresponding to the validation set... Sort in ascending order, traverse the sorted candidate list, and sort by the span of each candidate. as a threshold Calculate the evaluation result of the model at this threshold. Select to make The optimal (highest F1 score) threshold is used as ;

[0037] S63. Valid Answer Filtering: For the candidate list corresponding to the test set, retain... The candidate span is used to determine the final event trigger words, event types, and event arguments, thus completing the chapter-level event extraction.

[0038] Preferably, the process of generating the embedded representation in S3 also includes a historical answer embedding fusion step, specifically: embedding and encoding the event-related answers (including historical trigger word candidates, historical event type candidates, and historical argument candidates) extracted from the model's history to obtain a historical answer embedding vector; concatenating this historical answer embedding vector with the initial semantic embedding vector H generated in S31 to form a semantic embedding vector that integrates historical information; and then inputting the fused semantic embedding vector into... The process is performed to generate a final embedded representation that includes historical contextual associations, thereby enhancing the model's ability to model the relationships between event elements across sentences and paragraphs in document-level texts.

[0039] A text-level event extraction system based on multi-turn machine reading comprehension, specifically including:

[0040] The template building module is used to design appropriate question templates for the three sub-tasks of the chapter-level event extraction task: event trigger word recognition, trigger word classification, and event argument extraction. The question templates incorporate semantic information of argument roles and event association features, and the event argument extraction template matches the corresponding question words according to the argument role type.

[0041] The sequence splicing module is used to splice the question template generated by the template building module with the text to be processed according to a preset format to form a model input sequence. The preset format includes a start indicator, a separator, and a fixed arrangement order of questions and text.

[0042] The embedding representation module is used to encode the input sequence output by the sequence concatenation module using a pre-trained language model to generate an initial semantic embedding vector; it is also used to process the initial semantic embedding vector through a bidirectional gated recurrent unit to capture the bidirectional dependencies of the text and output the final embedding representation; it is also used to embed and encode the event-related answers extracted from the model history to obtain the historical answer embedding vector, and concatenate it with the initial semantic embedding vector to form a semantic embedding vector that integrates historical information.

[0043] The answer location module uses two independent binary classifiers to predict the probability of each tag in the final embedding representation generated by the embedding representation module as an event trigger word, event type, and the start and end positions of the event argument, thus obtaining the probability distribution of candidate answer positions.

[0044] The training optimization module is used to calculate the loss value based on the candidate answer location probability distribution and true label obtained by the answer location module, and then minimizes the loss value through the optimizer to update the model parameters.

[0045] The dynamic reasoning module is used to determine the optimal probability threshold based on the validation set, calculate the no-answer score of the candidate span by combining the candidate answer position probability distribution obtained by the answer location module, and filter out the candidate spans whose no-answer scores meet the threshold conditions as the final event trigger words and event arguments, thus completing the chapter-level event extraction.

[0046] The technical effects and advantages of this invention are as follows:

[0047] (1) This invention classifies argument roles by type and matches corresponding interrogative words to generate multiple question templates. It associates semantically similar argument roles (such as the participants and the deployers in the event participant class) in different event types through interrogative words with unified semantic attributes, establishes a semantic sharing channel across event types, breaks the limitation of independent classification of similar roles in traditional methods, enables the model to efficiently reuse the semantic features of similar roles, significantly improves the recognition and generalization ability of similar argument roles, and adapts to the argument extraction needs under multiple event types.

[0048] (2) This invention generates candidate results through preliminary answer positioning, and then performs final screening through dynamic threshold reasoning, replacing the traditional rigid pipeline process. This mechanism cuts off the transmission chain of errors in the preceding steps to the subsequent steps. Even if there is a small deviation in the preliminary positioning, it can be corrected through contextual adaptive screening of dynamic thresholds, effectively avoiding the problem of trigger word recognition errors affecting argument extraction, and ensuring the stability of the overall accuracy of event extraction;

[0049] (3) The present invention is achieved through and The embedded representation scheme of this invention can capture the semantic associations across sentences and paragraphs in the text at the chapter level, and improve the recall ability of scattered arguments; then, through dynamic threshold reasoning, the screening criteria are adaptively adjusted according to the text context to avoid the problem of missed selection and misselection in scenarios with multiple overlapping events and missing arguments under fixed thresholds, and finally achieve a dynamic balance between event extraction recall and precision, which is suitable for the needs of long text processing at the chapter level.

[0050] This invention flexibly adjusts training parameters based on the characteristics of texts in different domains, adapting to military, emergency, and other scenarios without reconstructing the model. By combining historical answer embedding and fusion, the model can optimize subsequent reasoning using the results of previous extractions, strengthen the modeling ability of cross-sentence event element associations, further enhance the practicality of chapter-level event extraction, and meet the needs of structured event knowledge extraction in real-world scenarios. Attached Figure Description

[0051] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used together with the embodiments of the invention to explain the invention, but do not constitute a limitation thereof. In the drawings:

[0052] Figure 1This is a schematic flowchart of the present invention;

[0053] Figure 2 This is a diagram showing the correspondence between semantic terms and argument roles in this invention;

[0054] Figure 3 This is a diagram showing the correspondence between some argument roles and problem templates in this invention;

[0055] Figure 4 The following are charts showing the ablation experiment results of the dynamic threshold decoding module of this invention;

[0056] Figure 5 The following are charts showing the ablation experiment results for different types of problems in this invention;

[0057] Figure 6 This is a chart showing the experimental results of model comparison in the DuEE-FIN dataset of this invention;

[0058] Figure 7 The chart shows the experimental results of model comparison in the CLFG-EEDMN dataset of this invention. Detailed Implementation

[0059] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0060] This invention provides, for example Figures 1-7The method for text-level event extraction based on multi-turn machine reading comprehension includes the following steps: S1, Question Template Construction: For the three sub-tasks of text-level event extraction—event trigger word recognition, trigger word classification, and event argument extraction—adaptive question templates are designed for each. The question templates incorporate semantic information of argument roles and event association features, and the event argument extraction template matches corresponding interrogative words based on the argument role type. S2, Input Sequence Concatenation: The question templates constructed in S1 are concatenated with the text to be processed according to a preset format to form the model input sequence. The preset format includes a start indicator, a separator, and a fixed arrangement order of questions and text. S3, Embedded Representation Generation: The input sequence obtained in S2 is encoded using a pre-trained language model to generate an initial semantic embedding vector. Then, a bidirectional gated recurrent unit is used to process the initial semantic embedding vector, capturing the bidirectional dependencies in the text and outputting the final embedded representation. S4, Answer Matching and Localization: Two independent binary classifiers are used to match and locate the answer generated in S3. In the final embedded representation, each tag is predicted as the probability of the start and end positions of the event trigger word, event type, and event argument, thus obtaining the probability distribution of candidate answer positions; S5, Model Training and Optimization: Based on the probability distribution of candidate answer positions obtained in S4 and the true labels, the cross-entropy loss function is used to calculate the loss value, and the model parameters are updated by minimizing the loss value through the optimizer; S6, Dynamic Threshold Inference: Based on the validation set, the optimal probability threshold is determined, and the no-answer score of the candidate span is calculated by combining the probability distribution of candidate answer positions obtained in S4. Candidate spans with no-answer scores that meet the threshold conditions are selected as the final event trigger word and event argument, completing the chapter-level event extraction. Through the complete chapter-level event extraction process of constructing question templates → input sequence splicing → embedded representation generation → answer matching and positioning → model training and optimization → dynamic threshold inference, the system systematically solves the multi-dimensional problems of traditional event extraction methods, such as the lack of argument role modeling, insufficient chapter-level context capture, pipeline error propagation, and rigid fixed threshold screening. This process deeply integrates the multi-turn machine reading comprehension (MRC) paradigm with dynamic reasoning mechanisms, achieving end-to-end processing from text input to structured event output. It can explicitly incorporate argument role semantic information through question templates and adapt to complex scenarios of chapter-level texts with the help of dynamic thresholds. Ultimately, it achieves an overall improvement in the accuracy, generalization ability and scenario adaptability of chapter-level event extraction, providing a unified and efficient technical framework for the structuring of long text events.

[0061] Specifically, the process of constructing the question template in step S1 includes: S11, Trigger word recognition template construction: designing a general interrogative sentence template for locating potential event trigger words. The template is composed of interrogative words combined with the core expression of trigger words, used to guide the model to filter out continuous words that can serve as event trigger words from the text; S12, Trigger word classification template construction: designing an interrogative sentence template containing the located potential trigger words, the location information of the trigger words in the text, and candidate event types. The template is constructed with confirmatory expressions, used to guide the model to judge the matching relationship between potential trigger words and candidate event types, and then determine the event type corresponding to the trigger words; S13, Event argument extraction template construction: first classifying the event argument role types into categories such as event participants, event location attributes, event time attributes, event equipment or scale or content attributes; for different categories of argument roles, matching the corresponding interrogative words: if the argument role For event participants, interrogative words pointing to the subject are selected; for event location attributes, interrogative words pointing to spatial location are selected; for event time attributes, interrogative words pointing to time range are selected; for event equipment, scale, or content attributes, interrogative words pointing to things or states are selected. Based on the matched interrogative words and argument role names, three types of event argument extraction templates are generated: the first type is a template containing only argument role names, the second type is a template containing the corresponding interrogative words and argument role names, and the third type is a template manually generated in the form of natural language, combined with the context of the text, to adapt to the needs of argument extraction with different semantic complexities. By designing special problem templates for the three sub-tasks of trigger word recognition, trigger word classification, and event argument extraction, and matching interrogative words according to argument role type to generate three types of templates in the event argument extraction templates, the problems of insufficient template semantic information and lack of modeling of similar argument roles in traditional MRC event extraction methods are solved. Among them, the trigger word recognition template can accurately guide the model to locate potential trigger words, the trigger word classification template can clarify the matching relationship between trigger words and event types, and the argument extraction template establishes semantic association of argument roles across event types through semantic classification and interrogative word matching, realizes the sharing of semantic features of similar argument roles, significantly improves the model's accuracy in recognizing argument roles under different event types, and is especially suitable for the extraction needs of specialized argument roles in fields such as military.

[0062] The following describes the construction of the trigger word identification, trigger word classification, and event argument extraction question templates. The trigger word identification stage aims to locate keywords that can trigger events. A general MRC question template is designed for this purpose, as shown in the example below: Q: Which word is the trigger word? Through this question, the model can extract consecutive words from a sentence or paragraph as potential trigger words. The event classification task, after identifying the trigger words, categorizes them into specific event types. For this purpose, the following question template is designed: Q: Is the trigger word <trigger word><trigger word position>, and the event type <event type>? Here, <trigger word> represents the identified trigger word, <trigger word position> indicates the position of the trigger word in the original text, and <event type> represents the candidate event type. This method is used to extract the answer to the question template. Traditional event trigger word identification and trigger word classification tasks typically rely on sequence labeling techniques to identify and classify trigger words. However, these methods may perform poorly when data is insufficient or when encountering new event types. In contrast, the MRC-based event detection task not only enhances the model's generalization ability but also improves its adaptability to new event types. Problems that fully incorporate prior information facilitate the learning of more semantic information. Before defining the problem template, we first analyze the constructed CLFG-EEDMN dataset (a dataset specifically built for event extraction tasks in the military news domain. EEDMN is an abbreviation for Event Extraction in Defense Military News, indicating its application domain and task direction. In the research on text-level event extraction based on multi-turn machine reading comprehension, this dataset is used to train and evaluate model performance, verify the effectiveness of the model in military news text processing, and help improve the model's ability to extract events in the military domain). In each event type framework, we found potential participating roles in each event type argument. There are two main types of arguments in the event arguments: event participants and event attributes. Event participants are labelable entity information that participates in an event in a certain type of event in a certain way. For example, <exercise participants> in the exercise event <exercise>; <supplyer> and <supply recipient> in the supply event <supply>. Event attributes are attributes related to the event other than participants, such as <time> and <location>.

[0063] Based on the above analysis, if the argument character is an event participant, the question uses the interrogative word "who"; if the argument character is the event-related attribute "location," the question uses the interrogative word "where"; if the character type is the event participant's related equipment, the question uses the interrogative word "what." The specific correspondences are as follows: Figure 2 As shown.

[0064] After identifying the semantic terms in the question templates, three types of templates were designed to explore the impact of different question template results on event extraction:

[0065] The first type of argument role extraction question template creates a template solely based on argument role names (e.g., duration, participants, location). The second type uses semantic terms and argument role names as questions, first determining the general semantic type of the argument role and then adding the argument role. The third type, to incorporate more semantic information and make the question description more natural, uses a manual method to generate questions for each argument role based on the event type framework in the CLFG-EEDMN dataset. The three question templates are as follows: Figure 3 As shown.

[0066] The specific method for concatenating the input sequence in step S2 is as follows: according to The pieces are assembled in the following order, among which... As the start indicator, Used as a separator; Let m be a sequence of problems. , The i-th word in the question; Let n be a sequence of text passages of length n. , Let j be the j-th word in the text; the total length of the concatenated input sequence is n+m+3 tokens, which is obtained by... This fixed-format concatenation of question templates and text addresses the problems of semantic separation between questions and text in traditional text input methods, and the difficulty for models to simultaneously capture question intent and textual information. This concatenation method clearly defines the position and function of start indicators and separators, ensuring that the input sequence contains both the task guidance information of the question and the complete context of the text. This guarantees that the pre-trained language model can simultaneously understand what needs to be extracted and from where, providing structurally clear and semantically relevant input data for the subsequent embedding representation generation stage, thus achieving effective fusion of question and text semantics.

[0067] The specific process of embedding representation generation in step S3 includes: S31, pre-trained language model encoding: inputting the input sequence concatenated in S2 into the pre-trained language model. (BidirectionalEncoderRepresentationsfromTransformers is a pre-trained language model proposed by Google in 2018), outputting the initial semantic embedding vector. , Where l is the length of the input sequence, and d is... The hidden layer dimension is R, which is the set of real numbers and satisfies the formula. (For illustration) Source: By concatenating the input sequence Send in The model is encoded, and the final output is the initial semantic embedding vector. S32, Processing: The initial semantic embedding vector obtained in S31 Input to bidirectional gated loop unit By employing a bidirectional gating mechanism to capture the positive and negative dependencies in the text, an enhanced contextual feature vector C is output. ,in S33, Probability Distribution Calculation: Perform linear transformations on the context feature vector C obtained in S32, and introduce the starting position weight matrix. End position weight matrix and the corresponding bias terms , Then, the probability of each marker being the start and end position is calculated using the Softmax function, satisfying the formula. , ,in Let be the probability of using the i-th marker as the starting position. Let the probability of the i-th mark being the ending position be given; thus, the probability distribution of the starting position can be obtained. Probability distribution of the ending position And the probability distribution of the starting position Probability distribution of the ending position Used for subsequent answer matching and dynamic threshold inference steps, through Encoding → The process of processing and calculating probability distributions to generate embedded representations solves the problems of traditional models' insufficient capture of bidirectional dependencies in document-level text and weak semantic representation capabilities. Among these, The model can extract deep contextual semantics and generate initial semantic embedding vectors; By employing a bidirectional gating mechanism, positive and negative dependencies in the text are further captured, strengthening cross-sentence semantic connections. Then, linear transformation and Softmax calculations are used to obtain the probability distributions of the start and end positions, providing accurate semantic support for subsequent answer location. This process achieves full mining of textual semantic information at the document level, effectively improving the model's ability to locate event elements in long texts.

[0068] The specific process of answer matching and location in step S4 includes: S41, preliminary location of trigger words: based on the probability distribution of the starting position output in S3. Probability distribution of the ending position S41: Select consecutive labeled sequences with probability values ​​greater than a preset initial threshold as potential event trigger word candidates for further filtering by the dynamic threshold inference module; S42: Preliminary event type matching: Substitute the potential event trigger word candidates obtained in S41 and their positions in the text into the trigger word classification template constructed in S1, and combine... and Calculate the matching probability between the trigger word candidate and each candidate event type, and select the candidate event type with the highest matching probability as the event type corresponding to that trigger word candidate, forming a trigger word candidate-event type association pair for subsequent verification by the dynamic threshold inference module; S43, preliminary argument localization: For the trigger word candidate-event type association pair formed in S42, call the event argument extraction template for the corresponding event type constructed in S1, and extract arguments based on the question words and argument role information in the template. and The process selects consecutive labeled sequences that match semantic roles as event argument candidates for final selection by the dynamic threshold inference module. Through a process of initial trigger word localization → initial event type matching → initial argument localization, this approach solves the problems of disconnect between event element localization and subsequent selection stages, and the resulting low quality of candidate results in traditional pipelined methods. Based on the probability distribution generated by embedded representations, this process first selects potential trigger words, trigger word candidate-event type association pairs, and argument candidates that meet the initial threshold. Each candidate is explicitly provided for further selection by the dynamic threshold inference module, providing a precise candidate basis for subsequent stages while avoiding errors caused by direct output results. This ensures the accuracy of initial event element localization and seamless integration with subsequent selection.

[0069] The specific process of model training and optimization in step S5 includes: S51, loss calculation: calculating the initial position loss using the cross-entropy loss function. Loss at the end position Satisfying the formula , ,in The starting position is the actual label sequence. Set the actual label sequence at the end position; set hyperparameters ∈[0,1], calculate the total loss And satisfy the formula S52, Parameter Update: Using the AdamW optimizer, the total loss is calculated based on S51. This method optimizes model training by updating the parameters of the pre-trained language model, bidirectional gated recurrent unit, and binary classifier through backpropagation. The optimization process, involving cross-entropy loss calculation followed by AdamW optimizer parameter updates, addresses the problems of single loss calculation, unstable parameter updates, and overfitting in traditional model training. Specifically, calculating the losses at the start and end positions separately and fusing the total loss with the hyperparameter α accurately optimizes the model's ability to predict answer boundaries. The AdamW optimizer, combined with weight decay characteristics, minimizes loss while suppressing overfitting, ensuring stable parameter updates. This process achieves both accuracy and stability in model training, ensuring rapid convergence to the optimal state in document-level event extraction tasks.

[0070] The parameter configuration and execution steps for model training include: S81, setting the hidden layer dimension: initializing the hidden layer dimension of the pre-trained language model and the bidirectional gated recurrent unit, setting it to a dimension value with sufficient semantic representation capability to ensure that the model can capture complex semantic relationships in document-level text; S82, configuring batch processing parameters: first setting a reasonable batch size, then adopting an accumulated step size strategy, that is, calculating the gradient for a preset number of consecutive small batches of training data, accumulating the gradients and performing a parameter update operation, simulating the effect of large batch training, reducing gradient oscillations in small batch training, and improving the stability of model training; S83, determining the training epochs and optimizer: setting the total number of model training epochs to the number of iterations that ensure the model can converge to the optimal parameter state; selecting AdamW as the model. The optimizer, leveraging its support for weight decay, regularizes and constrains model parameters to suppress overfitting. S84 sets learning rate and text length limits: the learning rate is set to balance learning speed and parameter update stability, preventing convergence due to excessively high learning rates or inefficient training due to excessively low learning rates. Simultaneously, the maximum length of the input text is limited, ensuring the model covers sufficient text information to handle long texts while controlling computational complexity, thus ensuring efficient training. This model training parameter configuration process—from hidden layer dimension setting to batch parameter configuration, training epoch and optimizer determination, and learning rate and text length limit setting—solves the problems of rigid model parameter settings, difficulty in adapting to different domain texts, and low training stability and efficiency associated with traditional methods. Among them, the higher-level parameter design (such as setting the hidden layer dimension with sufficient semantic representation ability and the learning rate that balances learning speed and stability) can be flexibly adjusted according to the characteristics of texts in different fields such as military and emergency response, without the need to reconstruct the model architecture; the gradient accumulation strategy can also improve the stability of mini-batch training, achieve a balance between model training efficiency and performance, and enhance the model's domain adaptability and engineering practicality.

[0071] The specific process of dynamic threshold inference in step S6 includes: S61, argument span collection: traversing all possible starting positions in the text. and end position ,in , ∈{1,...,M}, where M is the text length, filter spans that satisfy the following conditions: and Within the range of M, , and ,in , These are the start indicators. The corresponding starting and ending position probabilities; calculate the score for the span after filtering. Score for no answer ,Will Store in a candidate list, which includes potential event trigger word candidates (S41), trigger word candidate-event type association pairs (S42), and span information corresponding to event argument candidates (S43); S62, Determine the optimal threshold: Based on the candidate list corresponding to the validation set... Sort in ascending order, traverse the sorted candidate list, and sort by the span of each candidate. as a threshold Calculate the evaluation result of the model at this threshold. Select to make The optimal (highest F1 score) threshold is used as S63. Valid Answer Filtering: For the candidate list corresponding to the test set, retain... The process involves identifying candidate spans, determining the final event trigger words, event types, and event arguments, and completing text-level event extraction. A dynamic threshold reasoning process—argument span collection → optimal threshold determination → effective answer selection—solves the problem of traditional fixed threshold selection failing to adapt to complex text-level scenarios and prone to missed or incorrect selections. This process first collects candidate spans that meet the constraints and calculates scores for no-answer cases. Then, it optimizes the optimal threshold based on the validation set and finally selects valid answers that meet the threshold. It can adaptively adjust the selection criteria according to the text context (such as multiple overlapping events or cross-sentence arguments), effectively correcting deviations in the initial positioning stage and achieving accurate determination of the final event trigger words and arguments, balancing the recall and precision of event extraction.

[0072] Specifically, the S3 generation of embedded representations also includes a historical answer embedding fusion step, which involves: embedding and encoding the event-related answers (including historical trigger word candidates, historical event type candidates, and historical argument candidates) extracted from the model's history to obtain a historical answer embedding vector; concatenating this historical answer embedding vector with the initial semantic embedding vector H generated in S31 to form a semantic embedding vector that integrates historical information; and then inputting the fused semantic embedding vector into... The process involves processing to generate a final embedding representation that incorporates historical contextual relationships. This enhances the model's ability to model cross-sentence and cross-paragraph event element relationships within document-level texts. By integrating a historical answer embedding fusion step into the embedding representation generation process, the problem of traditional models being unable to utilize earlier extraction results and insufficiently modeling cross-sentence event element relationships in document-level texts is addressed. This step concatenates the embedding vectors of historical trigger word candidates, event type candidates, and argument candidates with the initial semantic embedding vector, and then inputs them... This processing enables the model to capture the semantic relationships between previous and current event elements, especially suitable for scenarios where arguments are distributed across sentences in text-level texts, achieving effective modeling of cross-sentence event element relationships and further improving the accuracy of argument extraction.

[0073] To evaluate the impact of the dynamic threshold decoding module on event extraction, an ablation experiment was designed. The specific impact of the module on event extraction was analyzed by comparing the results without the dynamic threshold decoding module (fixed probability threshold) and with the module. Experimental details are as follows. Figure 4 As shown:

[0074] In the experiment, the method without dynamic thresholding used a fixed probability threshold to filter valid arguments; that is, a fixed threshold was set, and any candidate segment with a probability greater than that threshold was considered a valid argument. This method failed to consider the potential impact of dynamically adjusting the threshold on different argument roles. The dynamic thresholding method calculates the optimal threshold on the validation set. And combine the unanswered scores for each character. Dynamic thresholding is used to dynamically filter candidate arguments, resulting in more accurate extraction. The dynamic thresholding method improves both P-value and F1 score, indicating that it can flexibly adjust the threshold in different scenarios, avoiding the extraction errors associated with fixed thresholds. In contrast, the accuracy of methods without dynamic thresholding is lower, suggesting that fixed thresholds fail to effectively handle the complexity of different argument types, leading to more extraction errors and inefficient candidate segment selection.

[0075] Three different argument extraction question templates were constructed. This invention section uses ablation experiments to investigate the impact of different question templates on argument extraction performance. The experimental comparison results are as follows: Figure 5 As shown:

[0076] The ablation experiment selected three different question template construction strategies. Template 1 simply constructs questions by directly using argument role names (such as duration, participants, etc.), which is simple in design but has limited information content. Template 2 further introduces the general semantic type of argument roles and constructs questions in the form of "is <semantic word> <argument role>?", aiming to inject preliminary semantic information into the questions. Template 3 uses the event type framework in the CLFG-EEDMN dataset to manually generate more natural and context-appropriate question descriptions, in order to incorporate more semantic information.

[0077] Experimental results show that Template 2, by introducing semantic type information, improves the recognition effect to some extent, but its overall performance improvement is limited due to the still limited information expressed in the questions. Template 3, through artificially generated question templates, contains richer semantic and argument role information, improving the accuracy and recall of argument recognition. Ablation experiments show that the design of the question template affects the performance of the event extraction task, especially since Template 3 encodes more semantic information about argument role names. In the application scenario of military news, the method of artificially constructing question templates by introducing argument role information can improve the accuracy of the event extraction model.

[0078] To verify the effectiveness of event extraction, this invention compares the proposed model with other event extraction models from different perspectives. These models include machine learning models, traditional BIO-based event extraction models, and traditional MRC event extraction models. Specific model descriptions are as follows:

[0079] (1) Machine learning model: MaxEnt adopts a machine learning method based on the maximum entropy model. The MaxEnt model is the earliest application of machine learning in event extraction.

[0080] (2) Traditional Neural Network Models: Deep learning-based event extraction models use convolutional neural networks (CNN), recurrent neural networks (RNN), or long short-term memory networks (LSTM) to encode text, learn text features, and further mine implicit semantic features in the text. CNN-BiLSTM combines convolutional neural networks with bidirectional LSTM networks for event extraction. The hybrid neural network approach can effectively extract word-level and sentence-level features.

[0081] (3) Pre-trained language model: This type of model adapts the pre-trained language model to document-level event extraction. BERT-Pipeline and ERNIE-Pipeline are based on BERT and ERNIE respectively, and use text classification and BIO methods to extract events.

[0082] (4) MRC-based event extraction method: The MRC-based event extraction paradigm models the event extraction task as a reading comprehension task. The Bert-QA model uses a pipeline approach for event detection subtasks and event argument extraction. The model sets different question templates for arguments in different event types, and extracts important information about the event from the news text through the question templates.

[0083] The experimental results comparing the MRC-EE model proposed in this invention with other mainstream event extraction methods on the DuEE-FIN public dataset are as follows: Figure 6 As shown.

[0084] Figure 6 The results of comparative experiments between the MRC-EE model and the baseline model on the DuEE-fin dataset are presented. The experiments show that the proposed MRC-EE model achieves F1 scores of 84.1% and 74.37% in event detection and argument extraction tasks, respectively.

[0085] MaxEnt, limited by its feature modeling capabilities, achieves an F1 score of only 50.6% for argument extraction. CNN-BiLSTM extracts word-level features through a hybrid network, but its sequence labeling mechanism suffers from ambiguous boundary delineation for long entities, resulting in an F1 score of 63.6%. While the pre-trained model BERT / ERNIE-pipeline offers stronger semantic representation capabilities compared to CNN-BiLSTM, it still suffers from error propagation issues. Bert-QA, based on MRC, employs a single-turn question-answering mechanism, but its question templates lack explicit modeling of argument roles, leading to insufficient recall for complex events in the military domain (such as exercise objectives linked across sentences). In contrast, MRC-EE, through a multi-turn dynamic question-answering mechanism, injects argument roles (such as the reconnaissance party) and trigger word information into the template. It addresses the ambiguous boundary delineation of long entities in military texts with a dynamic threshold span selector for context-aware boundary adjustment, achieving a 3.27% improvement in F1 score for argument extraction compared to the Bert-QA model. Experiments validate the effectiveness of the proposed MRC-EE model.

[0086] The experimental results comparing the MRC-EE model proposed in this invention with other mainstream event extraction methods on the CLFG-EEDMN dataset are as follows: Figure 7 As shown.

[0087] Figure 7Experimental results for the MRC-EE model and other event extraction models are presented. The MRC-EE model significantly improves experimental results on the CLFG-EEDMN dataset. This performance improvement is primarily due to the rich contextual interaction achieved by embedding argument role information in the question template, and secondly, because the dynamic threshold-based model decoding method effectively addresses long and nested entities. On the CLFG-EEDMN dataset, the F1 scores for the event detection subtask and event argument extraction subtask reached 79.4% and 69.4%, respectively.

[0088] The CNN-BiLSTM-based model extracts local features through a convolutional neural network (CNN) and combines it with a bidirectional LSTM to capture long-distance dependencies, extracting event arguments using sequence labeling. However, this method relies on manually generated feature templates and has weak modeling capabilities for complex semantics (nested entities, cross-sentence associations), resulting in a high error rate in recognizing long entity boundaries. The MRC-EE model improves the F1 score by 9.4% compared to the CNN-BiLSTM model, mainly due to the adaptive adjustment of the dynamic threshold to entity boundaries. The Bert-Pipeline uses a BERT pre-trained model for text encoding, completing trigger word detection and argument classification in two stages. The pipeline design leads to error propagation; if trigger word recognition is incorrect, it directly affects the accuracy of subsequent argument extraction. It lacks support for long entities, requiring rule processing after decoding. The MRC-EE model improves the P-value by 7.4% compared to the Bert-Pipeline by explicitly integrating argument roles into the question template and using a dynamic threshold strategy. The ERNIE-pipeline incorporates entity knowledge on top of BERT and employs a two-stage pipeline extraction. While entity knowledge injection improves the accuracy of entity recognition, it lacks sufficient modeling of event argument roles, and the static threshold setting makes long entities (such as an international waterway in a disputed sea area) easy to be segmented into multiple fragments.

[0089] The Bert-QA model employs a single-round machine reading comprehension approach, where the question template construction process only includes trigger words or event type information, lacking explicit modeling of event arguments. This simple question template construction is ill-suited to the complex event structures found in military news. Experimental results show that MRC-EE, through multi-round MRC dynamic question templates, incorporates not only event type features but also argument role information during the question template construction stage, giving the argument extraction task an advantage in handling scenarios with cross-sentence related arguments and long entities. The F1 score for the argument extraction task is approximately 3.5% higher than Bert-QA. The experimental results demonstrate the advantages of the proposed MRC-EE model compared to other models.

[0090] A text-level event extraction system based on multi-turn machine reading comprehension includes: a template construction module, which designs suitable question templates for the three sub-tasks of text-level event extraction: event trigger word recognition, trigger word classification, and event argument extraction. The question templates incorporate semantic information of argument roles and event association features, and the event argument extraction template matches corresponding interrogative words based on the argument role type; a sequence concatenation module, which concatenates the question templates generated by the template construction module with the text to be processed according to a preset format to form a model input sequence. The preset format includes a start indicator, a separator, and a fixed arrangement order of questions and text; an embedding representation module, which encodes the input sequence output by the sequence concatenation module using a pre-trained language model to generate an initial semantic embedding vector; it also processes the initial semantic embedding vector through a bidirectional gated recurrent unit to capture bidirectional dependencies in the text and output the final embedding representation; and it further processes the model's historical data. The extracted event-related answers are embedded and encoded to obtain historical answer embedding vectors, which are then concatenated with the initial semantic embedding vectors to form semantic embedding vectors that integrate historical information. The answer localization module uses two independent binary classifiers to predict the probability of each tag in the final embedding representation generated by the embedding representation module as an event trigger word, event type, and the start and end positions of event arguments, thus obtaining the candidate answer position probability distribution. The training and optimization module calculates the loss value using the cross-entropy loss function based on the candidate answer position probability distribution obtained from the answer localization module and the ground truth labels, minimizes the loss value through an optimizer, and updates the model parameters. The dynamic inference module determines the optimal probability threshold based on the validation set, calculates the no-answer score for candidate spans based on the candidate answer position probability distribution obtained from the answer localization module, and selects candidate spans whose no-answer scores meet the threshold condition as the final event trigger words and event arguments, completing the chapter-level event extraction.

[0091] Finally, it should be noted that the above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A text-level event extraction method based on multi-turn machine reading comprehension, characterized in that, Includes the following steps: S1. Question Template Construction: For the three sub-tasks of the chapter-level event extraction task, namely event trigger word identification, trigger word classification and event argument extraction, an appropriate question template is designed for each. The question template incorporates the semantic information of argument roles and event association features, and the event argument extraction template matches the corresponding question words according to the argument role type. S2. Input sequence concatenation: The question template constructed in S1 is concatenated with the text to be processed according to a preset format to form a model input sequence. The preset format includes a start indicator, a separator, and a fixed arrangement order of questions and text. S3, Embedded Representation Generation: The input sequence obtained in S2 is encoded using a pre-trained language model to generate an initial semantic embedding vector; then, the initial semantic embedding vector is processed by a bidirectional gated recurrent unit to capture the bidirectional dependencies of the text and output the final embedded representation. S4. Answer matching and localization: Using two independent binary classifiers, the probability of each tag in the final embedded representation generated in S3 as the start and end positions of the event trigger word, event type, and event argument is predicted to obtain the probability distribution of the candidate answer position. S5. Model Training and Optimization: Based on the candidate answer position probability distribution and true label obtained in S4, the cross-entropy loss function is used to calculate the loss value, the optimizer is used to minimize the loss value, and the model parameters are updated. S6. Dynamic Threshold Reasoning: Determine the optimal probability threshold based on the validation set, calculate the no-answer score of the candidate span by combining the candidate answer position probability distribution obtained in S4, and select the candidate spans whose no-answer scores meet the threshold conditions as the final event trigger words and event arguments to complete the chapter-level event extraction.

2. The method for extracting passage-level events based on multi-turn machine reading comprehension according to claim 1, characterized in that, The specific process of constructing the problem template in step S1 includes: S11. Construction of Trigger Word Recognition Template: Design a general interrogative sentence template for locating potential event trigger words. The template is composed of interrogative words combined with the core expression of trigger words, and is used to guide the model to filter out continuous words that can be used as event trigger words from the text. S12. Construction of Trigger Word Classification Template: Design an interrogative template containing the located potential trigger words, the location information of the trigger words in the text, and the candidate event types. The template is constructed with a confirmatory expression to guide the model to judge the matching relationship between potential trigger words and candidate event types, and then determine the event type corresponding to the trigger words. S13, Construction of event argument extraction template.

3. The method for extracting passage-level events based on multi-turn machine reading comprehension according to claim 1, characterized in that, The specific method for concatenating the input sequence in step S2 is as follows: according to The pieces are assembled in the following order, among which... As the start indicator, Used as a separator; Let m be a sequence of problems. , The i-th word in the question; Let n be a sequence of text passages of length n. , Let j be the j-th word in the text; the total length of the concatenated input sequence is n+m+3 tokens.

4. The method for extracting passage-level events based on multi-turn machine reading comprehension according to claim 1, characterized in that, The specific process of embedding representation generation in step S3 includes: S31. Pre-trained language model encoding: Input the input sequence concatenated from S2 into the pre-trained language model. Output the initial semantic embedding vector , Where l is the length of the input sequence, and d is... The hidden layer dimension, R, is the set of real numbers, and satisfies the formula ; S32, Processing: The initial semantic embedding vector obtained in S31 Input to bidirectional gated loop unit By employing a bidirectional gating mechanism to capture the positive and negative dependencies in the text, an enhanced contextual feature vector C is output. ,in This is the enhanced feature vector corresponding to the i-th label in the input sequence; S33. Probability Distribution Calculation: Perform linear transformations on the context feature vector C obtained in S32, and introduce the initial position weight matrix. End position weight matrix and the corresponding bias terms , Then, the probability of each marker being a start position and an end position is calculated using the Softmax function; thus, the probability distributions of the start position and the end position are obtained, and the probability distribution of the start position is... Probability distribution of the ending position Used for subsequent answer matching and dynamic threshold reasoning steps.

5. The method for extracting passage-level events based on multi-turn machine reading comprehension according to claim 4, characterized in that, The specific process of answer matching and location in step S4 includes: S41. Preliminary trigger word localization: based on the probability distribution of the starting position output from S3. Probability distribution of the ending position The continuous marker sequences with probability values ​​greater than the preset initial threshold are selected as potential event trigger word candidates for subsequent screening by the dynamic threshold reasoning module. S42. Preliminary Event Type Matching: Substitute the potential event trigger word candidates obtained in S41 and their positions in the text into the trigger word classification template constructed in S1, and combine... and Calculate the matching probability between the trigger word candidate and each candidate event type, select the candidate event type with the highest matching probability as the event type corresponding to the trigger word candidate, and form a trigger word candidate-event type association pair for subsequent verification by the dynamic threshold inference module; S43. Preliminary Argument Localization: Based on the trigger word candidate-event type association pair formed in S42, the event argument extraction template for the corresponding event type constructed in S1 is called. According to the interrogative words and argument role information in the template, from... and Continuous marker sequences that match semantic roles are selected as event argument candidates for final selection by the dynamic threshold inference module.

6. The method for extracting passage-level events based on multi-turn machine reading comprehension according to claim 1, characterized in that, The specific process of model training and optimization in step S5 includes: S51. Loss Calculation: Calculate the initial position loss using the cross-entropy loss function. Loss at the end position Set hyperparameters ∈[0,1], calculate the total loss ; S52, Parameter Update: Using the AdamW optimizer, the total loss is calculated based on S51. The model training is optimized by updating the parameters of the pre-trained language model, bidirectional gated recurrent unit, and binary classifier in the model through backpropagation.

7. The method for extracting passage-level events based on multi-turn machine reading comprehension according to claim 6, characterized in that, The parameter configuration and execution steps for model training include: S81. Set the hidden layer dimension: Initialize the hidden layer dimension of the pre-trained language model and the bidirectional gated recurrent unit, and set it to a dimension value with sufficient semantic representation ability to ensure that the model can capture complex semantic relationships in the text. S82. Configure batch processing parameters: First, set a reasonable batch size, and then adopt the cumulative step size strategy, that is, calculate the gradient for a preset number of consecutive small batches of training data, accumulate the gradients and perform a parameter update operation. This method simulates the effect of large batch training, reduces gradient oscillation in small batch training, and improves the stability of model training. S83. Determine the number of training epochs and the optimizer: Set the total number of training epochs to the number of iterations that ensures the model can converge to the optimal parameter state; select AdamW as the model optimizer, and use the characteristic of the optimizer supporting weight decay to regularize the model parameters and suppress model overfitting. S84. Set learning rate and text length limits: Set the learning rate of the model training to a value that balances the learning speed and the stability of parameter updates, so as to avoid the model training from not converging due to the learning rate being too high or the training efficiency being low due to the learning rate being too low; at the same time, limit the maximum length of the input text, so as to ensure that the model covers enough text information to handle long texts of the chapter level, while controlling the computational complexity of the model and ensuring that the training process proceeds efficiently.

8. The method for extracting passage-level events based on multi-turn machine reading comprehension according to claim 1, characterized in that, The specific process of dynamic threshold inference in step S6 includes: S61. Argument Span Collection: Traversing all possible starting positions in the text. and end position ,in , ∈{1,...,M}, where M is the text length, filter spans that satisfy the following conditions: and Within the range of M, , and ,in , These are the start indicators. The corresponding starting and ending position probabilities; calculate the score for the span after filtering. Score for no answer ,Will Store in a candidate list, which includes potential event trigger word candidates in S41, trigger word candidate-event type association pairs in S42, and span information corresponding to event argument candidates in S43; S62. Determining the optimal threshold: Based on the candidate list corresponding to the validation set... Sort in ascending order, traverse the sorted candidate list, and sort by the span of each candidate. as a threshold Calculate the evaluation result of the model at this threshold. Select to make The optimal threshold is as ; S63. Valid Answer Filtering: For the candidate list corresponding to the test set, retain... The candidate span is used to determine the final event trigger words, event types, and event arguments, thus completing the chapter-level event extraction.

9. The method for extracting passage-level events based on multi-turn machine reading comprehension according to claim 1, characterized in that, The S3 process for generating embedded representations also includes a historical answer embedding fusion step, specifically: embedding and encoding the event-related answers extracted from the model's history to obtain a historical answer embedding vector; concatenating this historical answer embedding vector with the initial semantic embedding vector H generated in S31 to form a semantic embedding vector that fuses historical information; and then inputting the fused semantic embedding vector into... The process is performed to generate a final embedded representation that includes historical contextual associations, thereby enhancing the model's ability to model the relationships between event elements across sentences and paragraphs in document-level texts.

10. A text-level event extraction system based on multi-turn machine reading comprehension, characterized in that, The system includes the text-level event extraction method based on multi-turn machine reading comprehension as described in any one of claims 1-9, specifically including: The template building module is used to design appropriate question templates for the three sub-tasks of the chapter-level event extraction task: event trigger word recognition, trigger word classification, and event argument extraction. The question templates incorporate semantic information of argument roles and event association features, and the event argument extraction template matches the corresponding question words according to the argument role type. The sequence splicing module is used to splice the question template generated by the template building module with the text to be processed according to a preset format to form a model input sequence. The preset format includes a start indicator, a separator, and a fixed arrangement order of questions and text. The embedding representation module is used to encode the input sequence output by the sequence concatenation module using a pre-trained language model to generate an initial semantic embedding vector; it is also used to process the initial semantic embedding vector through a bidirectional gated recurrent unit to capture the bidirectional dependencies of the text and output the final embedding representation; it is also used to embed and encode the event-related answers extracted from the model history to obtain the historical answer embedding vector, and concatenate it with the initial semantic embedding vector to form a semantic embedding vector that integrates historical information. The answer location module uses two independent binary classifiers to predict the probability of each tag in the final embedding representation generated by the embedding representation module as an event trigger word, event type, and the start and end positions of the event argument, thus obtaining the probability distribution of candidate answer positions. The training optimization module is used to calculate the loss value based on the candidate answer location probability distribution and true label obtained by the answer location module, and then minimizes the loss value through the optimizer to update the model parameters. The dynamic reasoning module is used to determine the optimal probability threshold based on the validation set, calculate the no-answer score of the candidate span by combining the candidate answer position probability distribution obtained by the answer location module, and filter out the candidate spans whose no-answer scores meet the threshold conditions as the final event trigger words and event arguments, thus completing the chapter-level event extraction.

Images